|Publication number||US7000048 B2|
|Application number||US 10/740,771|
|Publication date||Feb 14, 2006|
|Filing date||Dec 18, 2003|
|Priority date||Dec 18, 2003|
|Also published as||US20050138622|
|Publication number||10740771, 740771, US 7000048 B2, US 7000048B2, US-B2-7000048, US7000048 B2, US7000048B2|
|Inventors||Gary L. McAlpine, Dave B. Minturn|
|Original Assignee||Intel Corporation|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (16), Non-Patent Citations (3), Referenced by (46), Classifications (11), Legal Events (3)|
|External Links: USPTO, USPTO Assignment, Espacenet|
The invention relates generally to networking and, more particularly, to an apparatus and method for processing network data in parallel on a single execution thread of a processing device.
Computer networks, as well as other communication networks, routinely exchange information in units—commonly referred to as packets—corresponding to a known format. For example, a network may exchange packets associated with the TCP/IP protocols. See, e.g., Internet Engineering Task Force Request for Comment (IETF RFC) 791, Internet Protocol, and IETF RFC 793, Transmission Control Protocol. In addition to packets, other network data may also be exchanged over a computer network, including host requests/descriptors, timeout requests, transmission requests, as well as others.
Generally, operations performed on network data include receive processing, transmit processing, and timer processing. For example, when a packet is received at an end node of a network, such as a server or server cluster (or an intermediate device such as a router), one or more operations are usually performed with respect to the received packet. This packet processing may include, by way of example, accessing source and destination addresses and/or port numbers in the header of a received packet to make classification and/or routing decisions, flow control operations (e.g., sending a TCP ACK packet, error detection, etc.), queuing, as well as procedures associated with establishing or tearing down a connection.
Traditionally, processing of packets and other network data was performed on a general purpose processor supporting a single execution thread. Often times, the processing of a packet requires a large number of memory accesses to system memory or other off-chip memory. The single threaded processor performs operations sequentially and, therefore, it can stall during memory accesses and other slow operations while waiting for such operations to complete. Each stall in the processing of a packet due to a memory access (or other relatively slow process) wastes a significant number of clock cycles. The combination of a large number of unutilized clock cycles with the sequential nature of a single threaded processor creates an inefficient scheme for handling packets and other network data.
Processor clock cycles are shrinking at a much greater rate than memory access latencies. Thus, the number of clock cycles that may be wasted during a stall (e.g., for a memory access) is rapidly increasing, which in turn has caused the execution of packet processing—as well as the processing of other network data—on general purpose CPUs (central processing units) to rapidly decrease in efficiency, resulting in a failure to utilize the computational potential offered by high frequency processing devices. As equipment vendors strive to increase the speed and performance of general purpose computers (e.g., a client or server), the effects of the above-described failure to harness the abilities of high speed processors is becoming more profound.
The traditional method of using caches to reduce the frequency of memory accesses in application code is not very effective for processing packets and other network data due to a very low re-use of cached parameters and data. Also, conventional software multi-threading schemes do not provide a viable solution for the processing of network data. Today's multitasking operating systems (OS) may utilize methods of software multi-threading to share one or more processor execution threads between the many programs that may be executing on a computer simultaneously. However, OS multi-threading exhibits a very high overhead (e.g., thousands of clock cycles) because the OS implements a software scheduler common to all programs running on a system and, therefore, the OS has to deal not only with switching between threads but also with swapping program operating environments and contexts in and out of the CPU hardware in order to support the threads. Thus, due to these high thread switching latencies that typically consume thousands of clock cycles on conventional CPUs, such software multi-threading schemes cannot be utilized to hide memory accesses and other stalls that typically consume a few hundred clock cycles.
The masking of memory access latencies (and other stalls) experienced while processing packets and other network data may be achieved using multi-threaded hardware. Generally, hardware multi-threading schemes replicate certain hardware resources to facilitate parallel streams of execution. The use of multi-threaded processing hardware has been shown to be an effective method of hiding stalls for memory accesses and other slow operations. However, use of multi-threaded processors adds significant hardware complexity, while also increasing cost, real estate, and power consumption of the processing system. At the same time, these multi-threaded processors only provide a significant performance advantage for the few applications and operations that do not achieve effective use of cache. In addition, today's high volume (both in terms of manufacture and use) general purpose processors—e.g., those used for desk top and lap top computers—obtain less advantage from multi-threading than lower volume processors—e.g., those used for servers and workstations. Furthermore, it should be noted that, from a cost and power consumption standpoint, the use of lower complexity, higher volume, single threaded general purpose processing devices for desk top, lap top, server, work station, and packet processing alike, is desirable.
In the arena of network data processing, the use of a specialized packet processor supporting multiple threads of execution has been proposed. A multi-threaded packet processor can process multiple packets in parallel and very effectively reduce the performance cost of memory access stalls. However, these specialized packet processors suffer from many of the above-described disadvantages. Costs are increased due to added hardware complexity and lower volume markets. Further, in comparison to single threaded high volume processors (such as those used in desk top and lap top computers), the multi-threaded packet processor will have higher power requirements, increased cooling loads, and may be more difficult to program.
Disclosed are embodiments of a method and apparatus for processing network data, such as packets, in parallel on a single thread of a processing device. The disclosed embodiments are described below in the context of processing multiple packets in parallel on a single processing thread. However, it should be understood that the disclosed embodiments are applicable to the processing of any type of network data, whether packetized data or other type of data (e.g., host requests/descriptors, timeout requests, transmission requests, etc.) and, further, irrespective of whether the data processing is receive, transmit, or timer processing. As used herein, a “packet” is any network communication having a header and a payload (e.g., a TCP/IP packet). The disclosed embodiments are also described below in the context of an end node (e.g., a server) on a network. However, it should be understood that the disclosed embodiments are applicable to any device on a communications network (e.g., a router) that receives and handles network data.
Each of the nodes 110, 200 is coupled with a router 130 via a link 120. Node 200 is coupled with router 130 via link 120 k, whereas node 110 a is coupled with router 130 via link 120 a, node 110 b is coupled with router 130 via link 120 b, and so on. Each of the links 120 a–k may be established over any suitable medium—e.g., wireless, copper wire, fiber optic, or a combination thereof—supporting the exchange of information via any suitable protocol—e.g., TCP/IP (Transmission Control Protocol/Internet Protocol), HTTP (Hyper-Text Transmission Protocol), as well as others.
The router 130 couples the network 100 with another network (or networks) 5, such as, by way of example, the Internet and/or another LAN, MAN, WAN, or WLAN. The router 130 comprises any device (or combination of devices) that can receive from the other network 5 packets (or other network data) that are destined for the network 100 and route these packets to the appropriate destination within network 100. Router 130 also receives packets (or other network data) from the nodes 110, 200 within the network 100 and forwards these communications to the other network 5. The router 130 may be coupled with the other network 5 via any suitable medium, including a wireless, copper wire, and/or fiber optic connection using any suitable protocol (e.g., TCP/IP, HTTP, etc.).
It should be understood that the network 100 shown in
As noted above, at least the node 200 implements an embodiment of a disclosed parallel processing agent, as will be described below. In one embodiment, the node 200 comprises any suitable computer system (e.g., a server), and the parallel processing agent comprises a software application that may be implemented or executed on this computer system. An embodiment of such a computer system is illustrated in
Coupled with bus 205 is a processing system 300. The processing system 300 may comprise any suitable processing device or combination of processing devices, and embodiments of the processing system 300 are illustrated below in each of
Computer system 200 also includes system memory 210 coupled with bus 205. The system memory 210 may comprise any suitable type and number of random access memories, such as static random access memory (SRAM), a dynamic random access memory (DRAM), a synchronous DRAM (SDRAM), and/or a double data rate DRAM (DDRDRAM). During operation of computer system 200, an operating system 214, a parallel processing agent 400, as well as other programs 218 may be resident in the system memory 210. Computer system 200 may further include a read-only memory (ROM) 220 coupled with the bus 205. During operation, the ROM 220 may store instructions and variables for processing system 300, and ROM 220 may also have resident thereon a system BIOS (Basic Input/Output System). The computer system 200 may also include a storage device 230 coupled with the bus 205. The storage device 230 comprises any suitable non-volatile memory, such as, for example, a hard disk drive. The parallel processing agent 400, as well as operating system 214 and other programs 218, may be stored in the storage device 230. Further, a device 240 for accessing removable storage media (e.g., a floppy disk drive or a CD ROM drive) may be coupled with bus 205.
The computer system 200 may include one or more input devices 250 coupled with the bus 205. Common input devices 250 include keyboards, pointing devices such as a mouse, and scanners or other data entry devices. One or more output devices 260 may also be coupled with the bus 205. Common output devices 260 include video monitors, printing devices, and audio output devices.
Computer system 200 further comprises a network interface 270 coupled with bus 205. The network interface 270 comprises any suitable hardware, software, or combination of hardware and software capable of coupling the computer system 200 with the network (or networks) 5 (and/or with router 130).
It should be understood that the computer system 200 illustrated in
In one embodiment, the parallel processing agent 400 comprises a set of instructions (e.g., a software application) run on a computer system—e.g., the computer system 200 of
In another embodiment, the parallel processing agent 400 comprises a set of instructions stored on a machine accessible medium, such as, for example, a magnetic media (e.g., a floppy disk or magnetic tape), an optically accessible disk (e.g., a CD-ROM disk), a flash memory device, etc. To run parallel processing agent 400 on, for example, computer system 200, the device 240 for accessing removable storage media may access the instructions on the machine accessible medium, and the instructions may then be executed in processing system 300. In this embodiment, the instructions (or a portion thereof) may again be downloaded to system memory 210.
In another embodiment, the parallel processing agent 400 is implemented in hardware or a combination of hardware and software (e.g., firmware). For example, the parallel processing agent 400 may be implemented in an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other similar device that has been programmed in accordance with the disclosed embodiments.
As previously noted, embodiments of the processing system 300 are illustrated in
Referring first to
The network data processing device 320 comprises any suitable processing device, such as a microprocessor, a network processor, an ASIC, an FPGA, or similar device. In one embodiment, the network processing device 320 comprises a general purpose processing system that supports a single thread 325, and the parallel processing agent 400 can be executed on this single thread. Although a single network data processing device 320 is illustrated in
As noted above, the network data processing device 320 may support a single thread of execution. A “thread” is a separate flow of control within a computer program. The thread shares the resources that are allocated to the program (e.g., address space, global variables, etc.) with other parts of—and/or other threads associated with—the program. The thread also has its own resources (e.g., execution stack, stack pointers, program counters, etc.) that are typically not shared with other components of the program. As used herein, the term “thread” encompasses a flow of control within a program and the resources allocated to that thread. Furthermore, a processing device is said to “support” a thread if that device includes at least some of the hardware (e.g., registers, memory space, as well as other circuitry) needed for the thread to utilize the resources available to it. In addition, the phrases “thread”, “execution thread”, “thread of execution”, “processing thread”, and the like are used herein interchangeably.
Referring next to
Turning now to
A core 310 and a processing engine 320 are coupled with the local bus 305. In one embodiment, the core 310 comprises a general purpose processing system, which may execute operating system 214. Core 310 may also control operation of processing system 300 and perform a variety of management functions, such as dispensing instructions to the processing engine 320 for execution. The processing engine 320 comprises any suitable processing system, and it may include an arithmetic and logic unit (ALU), a controller, and a number of registers (for storing data during read/write operations). The processing engine 320 supports a single thread of execution 325, and the parallel processing agent 400 may be executed on this thread. In another embodiment, the processing engine 320 supports multiple execution threads (e.g., four), and the parallel processing agent 400 may be executed on any one of these threads (the remaining threads being allocated, in one embodiment, to other tasks).
Also coupled with the local bus 305 is an on-chip memory subsystem 330. Although depicted as a single unit, it should be understood that the on-chip memory subsystem 330 may—and, in practice, likely does—comprise a number of distinct memory units and/or memory types. For example, such on-chip memory may include SDRAM, SRAM, and/or flash memory (e.g., FlashROM). It should be understood that, in addition to on-chip memory, the processing system 300 may be coupled with off-chip memory (e.g., ROM 220, off-chip cache memory, etc.).
Processing system 300 further includes a bus interface 340 coupled with local bus 305. Bus interface 340 provides an interface with other components of computer system 200, including bus 205. For simplicity, bus interface 340 is depicted as a single functional unit; however, it should be understood that, in practice, the processing system 300 may include multiple bus interfaces. For example, the processing system 300 may include a PCI bus interface, an IX (Internet Exchange) bus interface, as well as others, and the bus interface 340 is intended to represent a collection of one or more such interfaces.
It should be understood that the embodiment of processing system 300 illustrated and described with respect to
Yet another embodiment of the processing system 300 is illustrated in
In the embodiments of
In yet another embodiment, as shown in
The parallel processing agent 400 decomposes the processing of packets into a number of processing streams—each processing stream being referred to herein as a “pipeline”—and execution of these separate processing streams is interleaved on a single thread of a processing device (e.g., the packet processing device 320 of
Because the disclosed parallel processing agent enables the parallel processing of network data on a single thread, the disclosed embodiments are amenable to a system architecture in which host CPU processing is partitioned from the processing of network data. An embodiment of such an architecture 290 for the computer system 200 of
The network data processing system carries out TCP/IP stack processing and the processing of other network data, where the network data processing system comprises, for example, the network data processing device 320 of
The architecture 290 also includes a kernel web server 294 and kernel space applications 295 a, both of which may exchange data with the DTIs 297. The kernel space applications 295 a may include its own socket I/O 295 b, as shown in
Turning now to
Referring next to
The entire set of possible states 430 a–x within the pipelines 420 a–n is generally the same within each pipeline (although it is within the scope of the disclosed embodiments that the set of possible states vary between pipelines). However, as previously suggested, a packet (or other network data) will progress through a subset of these states 430 a–x while being transferred between the host processor 310 and the network 5, or between the network and the host processor. The actual sequence of states 430 a–x for any given packet (or other instance of network data) is dependent upon decisions made during processing of that packet, and this subset of the states 430 a–x for a given packet will be dependent upon a variety of factors. For example, the subset of the states 430 a–x that a packet progresses through may depend on the direction of transfer (e.g., is the packet arriving from a host processor or from a network), the state of the connection over which the packet was or will be transmitted (e.g., is the connection an established connection, a connection in the process of being set up, or a connection being torn down), and the contents of the packet's header.
As noted above, each of the states comprises a sequence of one or more operations that can be performed with respect to a packet (or other instance of network data). When a state is executed for a packet, the processing of that packet progresses through the sequence of operations associated with that state to reach a specified outcome or decision point. The outcome or decision point reached during execution of any given state generally coincides with a memory access or other operation that requires a relatively large number of clock cycles to complete, and processing of a packet on that pipeline cannot resume until the operation has been completed. Such events are referred to herein as “yield points” or “yield events.” When a pipeline reaches a yield point—which typically, but not necessarily, corresponds with the end of a state—the scheduler 410 will scheduler another one of the pipelines 420 for execution on the thread. Potential yield events include memory accesses (e.g., for non-cached data), the end of a network operation (e.g., the end of processing a packet), TCP context lookups and/or accesses, host interface interactions such as reading a queue, resource management tasks such as tracking counters and timers, asynchronous interactions with other components, and various hardware operations such as DMA (direct memory access) transfers, network interface interactions (e.g., descriptor, packet, and header fetches), virtual to physical address translation look-ups, etc. A yield event may also be an explicit yield performed by software (e.g., the parallel processing agent 400). For example, an explicit yield may be carried out to perform a speculative memory prefetch.
The number of pipelines 420 a–n needed for efficient operation is generally a function of the relationship between the number of clock cycles necessary for the scheduler to perform a pipeline switch (e.g., 28 clock cycles), the number of clock cycles needed to process the individual states, and the number of clock cycles necessary to perform a memory access or other asynchronous operation (e.g., 250 clock cycles). Generally, multiple pipelines will be needed to reduce or “hide” the effects of memory access latencies (or latencies associated with other operations) and to insure the number of unutilized clock cycles is minimized. In one embodiment, the number of pipelines comprises between two (2) and eight (8) pipelines; however, it should be understood that any suitable number of pipelines (e.g., >8) may be employed.
The scheduler 410 multiplexes the pipelines 420 a–n on the single thread being used by the parallel processing agent 400. The scheduler 410 includes a number of state pointers 440 (including state pointers 440 a, 440 b, . . . , 440 n) and a pipeline pointer 450, as well as sequence logic 460. Each of the state pointers 440 a–n corresponds to one of the pipelines 420 a–n (i.e., state pointer 440 a corresponds with pipeline 420 a, and so on), and the state pointer 440 associated with any given pipeline 420 keeps track of the current state within that pipeline. When one pipeline reaches a yield point, and the scheduler switches allocation of the thread to another pipeline, the state pointer 440 of the stalled pipeline will indicate which of the states 430 a–x is to be executed when that pipeline is resumed. In other words, the state pointer 440 for a pipeline 420 is a placeholder for the pipeline, such that the next state to execute on the pipeline is known when that pipeline is again allocated control of the thread.
The pipeline pointer 450 identifies the pipeline that currently has control of the thread upon which the parallel processing agent 400 is executing. When a pipeline 420 yields (e.g., when a state has completed), the sequence logic 460 will determine which of the other pipelines 420 is next to be allocated the thread. The sequence logic 460 may employ any suitable scheduling scheme, such as a round-robin scheduling scheme in which the pipelines 420 are sequenced for execution one by one in order, a scheme based on cache line locality, or a scheme based on a quantum. Further, the scheduling of the pipelines 420 may performed synchronously, asynchronously, or using some combination of both synchronous and asynchronous scheduling.
As noted above, the pipelines 420 may yield and be re-scheduled in a synchronous or asynchronous manner. For example, if a particular operation (e.g., a particular memory access) has a high probability of always causing a stall and the approximate length of the stall is always known, the pipelines may be configured to always yield at that point. Accordingly, a pipeline may be synchronously re-scheduled after a given time period, without regard for whether the stall event has completed. On the other hand, for operations where the probability of causing a stall is low or unknown and the length of the stall may vary widely or is unknown, the pipelines may be configured to either always yield or yield only if a stall event occurs. If a pipeline yields, it may then be asynchronously re-scheduled only after the stall event has completed. In one embodiment, rather than waiting a minimum time period for a memory access (or other operation), the scheduler 410 first checks for completion of the operation and, if not complete, the scheduler advances to the next pipeline. In another embodiment, if a pipeline is not be ready to commence execution when the thread has been allocated to it (e.g., because the memory access or other operation that earlier caused the pipeline to yield has not yet been completed), the pipeline may simply wait for the operation to complete while maintaining control of the thread (which, in a further embodiment, may occur after synchronous scheduling of the pipeline). Thus, it should be understood that the scheduler 410 and pipelines 420 may utilize any combination of synchronous and asynchronous yielding and re-scheduling methods, depending upon the nature of the stall events being avoided.
In another embodiment, as noted above, the sequence logic 460 may base scheduling decisions on cache line locality. In this scheme, the next pipeline 420 to be allocated the execution thread is determined according to which of the pipelines 420 a–n has the data it needs to resume execution in cache memory. The cache memory may comprise, for example, an on-chip memory of the network data processing device 320 of
In a further embodiment, which was also noted above, the sequence logic 460 may schedule pipelines on the basis of a quantum assigned to each pipeline 420. A quantum may be expressed as a certain quantity of resources (e.g., a specified number of clock cycles, a certain number of operations, etc.) allocated to a pipeline during any given opportunity to control the thread (each pipeline being given one opportunity to control the thread during a scheduling round). During any given scheduling round, a pipeline 420 cannot exceed its quantum of allocated resources. In this scheme, the quantum for each of the pipelines 420 a–n may be equal or, alternatively, the quantum may vary for the pipelines to allow some pipelines to obtain more resources than others (e.g., as part of a quality of service or resource reservation policy).
The parallel processing agent 400 also includes a context 470 associated with each of the pipelines 420 (e.g., context 470 a is associated with pipeline 420 a, and so on). The context 470 of each pipeline 420 stores state information regarding the packet (or other network data) currently being processed on that pipeline. The context 470 a–n of each of the pipelines 420 a–n, respectively, may be shared between the pipeline and the scheduler 410. When a pipeline 420 executes one of the states 430 a–x, the context 470 associated with that pipeline may be updated to reflect the outcome of that state. It should be understood that, although illustrated in
The data structures—e.g., state pointers 440 a–n, pipeline pointer 460, contexts 470 a–n—associated with the parallel processing agent 400 may be stored in any suitable memory and, in one embodiment, these data structures are stored in a memory directly accessible by the processing system. For example, these data structures (or a portion thereof) may be stored in on-chip memory, such as a SRAM, DRAM, SDRAM, DDRDRAM, or other memory local to the packet processing device 320 (see
As set forth in block 510 a, network data (e.g., a packet) is received in the first pipeline 505 a, and if the thread is allocated to the first pipeline—see block 520 a—the network data is processed on this pipeline, as set forth in block 530 a. Referring to block 540 a, if a yield event occurs in the first pipeline, control of the thread is returned to the scheduler, as set forth in block 550 a. The scheduler will then shift control of the thread to the next scheduled pipeline (e.g., one of pipelines 505 b through 505 n). With reference to block 560 a, if the processing of the network data is complete, the network data is disposed of, as set forth in block 570 a. For example, if the network data is a packet, the packet may be consumed or queued. If processing of the network data is not complete (see block 560 a), the first pipeline is idle until this pipeline is again allocated control of the thread (see block 520 a). New network data brought into the first pipeline (see block 510 a) is processed in a similar manner.
Each of the other processing streams 505 b through 505 n handles network data in a similar fashion. When any one of the processing streams 505 a–n yields, control of the thread is returned to the scheduler (see
Referring now to
As set forth in block 610, the current pipeline is set to the pipeline pointer (see
Referring to block 630, the next state for the current pipeline is determined. As noted above, the subset of states through which any network data progresses within a pipeline will depend on decisions made during processing of that data. Factors which can effect the next state of a pipeline (and the subset of states that the network data will progress through) include the direction of transfer of the network data (i.e., ingress vs. egress), the state of the connection over which the network data was or will be transmitted (i.e., established connection vs. a connection being set up or torn down), and the contents of the network data (e.g., a packet's header). The state pointer of the current pipeline is then incremented to the next state, which is set forth at block 635. When this pipeline again resumes control of the thread, the next state will then become the current state.
In one embodiment, prior to switching control of the thread to another pipeline, it is determined whether there is a cache hit for data needed by the current pipeline (i.e., data needed by the next state), as set forth in block 680. In one embodiment, a cache hit occurs when a memory access for the needed data can be completed from an on-chip memory. If there is a cache hit, the current state of the current pipeline is set to the state pointer (which was previously incremented), and this state is executed (see block 615 and 620). The context of the current pipeline is again updated (see block 625), and the next state is determined (see block 630) and the state pointer of the current pipeline incremented to the next state (see block 635).
If, however, there was not a cache hit (see block 680), the current pipeline will yield, as set forth in block 640. When the current pipeline yields, control of the thread is returned to the scheduler (see
Referring to block 655, if the end of processing the network data on the current pipeline has been reached, the network data is disposed of, as set forth at block 660. For example, if the network data is a packet, the packet may be consumed or queued. Upon disposition of the completed network data, another set of network data can be received in this pipeline (refer again to block 605). If the end of processing the network data has not been reached (or after disposing of the completed network data), the current pipeline is set—refer again to block 610—to the pipeline identified by the pipeline pointer (which was previously set to the next pipeline in block 650), and the above-described processes (i.e., blocks 615 through 680) is repeated for the pipeline that is now the current pipeline.
The above-described methods for parallel processing of network data on a single thread may be better understood with reference to the timing diagram 700 illustrated in
During a first time period 701, the scheduler 410 is performing a thread switch and determining which pipeline the thread should be allocated to. At a second time period 702, the first pipeline 420 a executes a state (i.e., a state “S1”). After the first pipeline completes execution of the state, control of the thread is again returned to the scheduler (for time period 703) to determine which pipeline to next schedule on the thread (or, in an alternative embodiment, during time period 703, the current pipeline determines the next pipeline using a shared schedule and transfers control of the thread directly to the next pipeline). Employing, for example, a round-robin scheduling scheme, the second pipeline 420 b is given control of the thread, and this pipeline executes a state (i.e., S3) during a fourth time period 704. During the next time period 705, control of the thread is returned to the scheduler to determine the next pipeline (or, according to the alternative embodiment, the current pipeline determines the next pipeline and transfers control of the thread directly to the next pipeline). During time period 706, the third pipeline 420 c executes a state (i.e., S1) and, after the scheduler (or current pipeline) performs pipeline switching during the next time period 707, the fourth pipeline 420 d executes a state (i.e., S2) during a subsequent time period 708, after which control of the thread is again returned to the scheduler for time period 709.
The above-described sequence is then repeated, wherein control of the thread is continually being multiplexed between the scheduler and the pipelines in a round-robin fashion (or according to a different scheduling scheme in other embodiments). It should be noted that the time periods 702, 704, 706, . . . , 716 during which the pipelines are executing a state are not necessarily equal, as the time required for execution may vary among the states. Also, as noted above, the states that a pipeline progresses through while processing any given network data will vary with the decisions made during processing of that data, and a given set of network data will typically progress through a subset of all possible states within a pipeline. This is also illustrated in
As described above with respect to
Embodiments of a parallel processing agent 400 providing multiple processing streams that are multiplexed on a single thread have been disclosed. However, it should be understood that the disclosed embodiments are not limited to the processing of network data. Rather, the disclosed embodiments for multiplexing multiple processing streams on a single thread may find application where it is desirable to process any type of data in parallel on a single thread of a processing device. Any application can utilize the disclosed pipeline/state structure with a dedicated scheduler in order to reduce sensitivity to memory access latencies and various other stall events.
Embodiments of a parallel processing agent capable of handling multiple processing streams in parallel on a single thread—as well as embodiments of a method of processing network data in parallel on a single thread—having been described above, those of ordinary skill in the art will appreciate the advantages of the disclosed embodiments. A pipeline can be entered at any appropriate state, executed on the thread until a yield event occurs, and then control of the thread returned to the scheduler to schedule the next pipeline (or transferred directly to the next pipeline). Thus, control of the thread can be continuously multiplexed amongst all pipelines. By multiplexing a single thread between a number of processing pipelines, clock cycles which might otherwise be wasted during a memory access are utilized for network data processing. Thus, greater efficiency and throughput are achieved. Further, because the disclosed parallel processing scheme can be executed on a single thread, the parallel processing agent is amenable to execution on a low cost and less power hungry general purpose processors that support a single thread of execution. Accordingly, efficient parallel processing is realized without the added cost and complexity of multi-threaded or multi-processor systems (although the disclosed parallel processing agent may, in some embodiments, be implemented on such systems).
The foregoing detailed description and accompanying drawings are only illustrative and not restrictive. They have been provided primarily for a clear and comprehensive understanding of the disclosed embodiments and no unnecessary limitations are to be understood therefrom. Numerous additions, deletions, and modifications to the embodiments described herein, as well as alternative arrangements, may be devised by those skilled in the art without departing from the spirit of the disclosed embodiments and the scope of the appended claims.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US4843540 *||Sep 2, 1986||Jun 27, 1989||The Trustees Of Columbia University In The City Of New York||Parallel processing method|
|US4974223||Sep 18, 1989||Nov 27, 1990||International Business Machines Corporation||Parallel architecture for high speed flag detection and packet identification|
|US5339415||Nov 9, 1992||Aug 16, 1994||Cray Research, Inc.||Dual level scheduling of processes to multiple parallel regions of a multi-threaded program on a tightly coupled multiprocessor computer system|
|US5357617||Nov 22, 1991||Oct 18, 1994||International Business Machines Corporation||Method and apparatus for substantially concurrent multiple instruction thread processing by a single pipeline processor|
|US5815727||Dec 20, 1995||Sep 29, 1998||Nec Corporation||Parallel processor for executing plural thread program in parallel using virtual thread numbers|
|US6079008||Apr 3, 1998||Jun 20, 2000||Patton Electronics Co.||Multiple thread multiple data predictive coded parallel processing system and method|
|US6085248 *||Apr 24, 1997||Jul 4, 2000||Xaqtu Corporation||Media access control transmitter and parallel network management system|
|US6415332||Aug 19, 1998||Jul 2, 2002||International Business Machines Corporation||Method for handling of asynchronous message packet in a multi-node threaded computing environment|
|US6516361 *||Sep 17, 1998||Feb 4, 2003||Sony Corporation||Method of and apparatus for capturing and processing continuous media-based data streams transmitted over an IEEE 1394 serial bus|
|US6532509||Dec 22, 1999||Mar 11, 2003||Intel Corporation||Arbitrating command requests in a parallel multi-threaded processing system|
|US6587906||Jan 9, 2003||Jul 1, 2003||Intel Corporation||Parallel multi-threaded processing|
|US6745237 *||Jan 15, 1998||Jun 1, 2004||Mci Communications Corporation||Method and apparatus for managing delivery of multimedia content in a communications system|
|US20020083297||Dec 22, 2000||Jun 27, 2002||Modelski Richard P.||Multi-thread packet processor|
|US20030014471||Apr 29, 2002||Jan 16, 2003||Nec Corporation||Multi-thread execution method and parallel processor system|
|US20030069920||Sep 24, 2002||Apr 10, 2003||Melvin Stephen W.||Multi-threaded packet processing engine for stateful packet processing|
|US20030081600||Oct 30, 2001||May 1, 2003||Blaker David M.||Methods, systems and computer program products for packet ordering for parallel packet transform processing|
|1||*||"A programming interface for network resource management" by Takahashi et al. (abstract only)□□Publication date: Mar. 26-27, 1999.|
|2||McAlpine, et al., "Embedded Transport Acceleration Architecture", U.S. Appl. No.: 10/305,738, Filed: Nov. 27, 2002, 41 Pgs.|
|3||Ungerer, et al., "A Survey of Processors With Explicit Multithreading", ACM Computing Surveys, vol. 35, No. 1, Mar. 2003, pp. 29-63.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US7441245 *||Aug 14, 2003||Oct 21, 2008||Intel Corporation||Phasing for a multi-threaded network processor|
|US7461213||Jul 23, 2004||Dec 2, 2008||Rmi Corporation||Advanced processor system using request, data, snoop, and response rings|
|US7509462||Aug 31, 2004||Mar 24, 2009||Rmi Corporation||Advanced processor with use of bridges on a data movement ring for optimal redirection of memory and I/O traffic|
|US7509476||Feb 8, 2007||Mar 24, 2009||Rmi Corporation||Advanced processor translation lookaside buffer management in a multithreaded system|
|US7627721||Jul 23, 2004||Dec 1, 2009||Rmi Corporation||Advanced processor with cache coherency|
|US7739436 *||Nov 1, 2004||Jun 15, 2010||Sonics, Inc.||Method and apparatus for round robin resource arbitration with a fast request to grant response|
|US7788448 *||Jan 19, 2007||Aug 31, 2010||Pmc-Sierra Us, Inc.||Sequencer cache and method for operating the same|
|US7924828||Aug 31, 2004||Apr 12, 2011||Netlogic Microsystems, Inc.||Advanced processor with mechanism for fast packet queuing operations|
|US7941603||Nov 30, 2009||May 10, 2011||Netlogic Microsystems, Inc.||Method and apparatus for implementing cache coherency of a processor|
|US7961723||Aug 31, 2004||Jun 14, 2011||Netlogic Microsystems, Inc.||Advanced processor with mechanism for enforcing ordering between information sent on two independent networks|
|US7984268||Jul 23, 2004||Jul 19, 2011||Netlogic Microsystems, Inc.||Advanced processor scheduling in a multithreaded system|
|US7991977||Dec 20, 2007||Aug 2, 2011||Netlogic Microsystems, Inc.||Advanced processor translation lookaside buffer management in a multithreaded system|
|US8015567||Aug 31, 2004||Sep 6, 2011||Netlogic Microsystems, Inc.||Advanced processor with mechanism for packet distribution at high line rate|
|US8037224||Jul 31, 2007||Oct 11, 2011||Netlogic Microsystems, Inc.||Delegating network processor operations to star topology serial bus interfaces|
|US8065456||Jan 24, 2008||Nov 22, 2011||Netlogic Microsystems, Inc.||Delegating network processor operations to star topology serial bus interfaces|
|US8176298||Aug 31, 2004||May 8, 2012||Netlogic Microsystems, Inc.||Multi-core multi-threaded processing systems with instruction reordering in an in-order pipeline|
|US8478811||Oct 30, 2008||Jul 2, 2013||Netlogic Microsystems, Inc.||Advanced processor with credit based scheme for optimal packet flow in a multi-processor system on a chip|
|US8499302||Sep 6, 2011||Jul 30, 2013||Netlogic Microsystems, Inc.||Advanced processor with mechanism for packet distribution at high line rate|
|US8543747||Oct 4, 2011||Sep 24, 2013||Netlogic Microsystems, Inc.||Delegating network processor operations to star topology serial bus interfaces|
|US8561076||Jun 30, 2004||Oct 15, 2013||Emc Corporation||Prioritization and queuing of media requests|
|US8788732||Aug 21, 2013||Jul 22, 2014||Netlogic Microsystems, Inc.||Messaging network for processing data using multiple processor cores|
|US8856452||May 31, 2011||Oct 7, 2014||Illinois Institute Of Technology||Timing-aware data prefetching for microprocessors|
|US8953628||Jun 6, 2011||Feb 10, 2015||Netlogic Microsystems, Inc.||Processor with packet ordering device|
|US9088474||Aug 31, 2004||Jul 21, 2015||Broadcom Corporation||Advanced processor with interfacing messaging network to a CPU|
|US9092360||Aug 1, 2011||Jul 28, 2015||Broadcom Corporation||Advanced processor translation lookaside buffer management in a multithreaded system|
|US9154443||Oct 20, 2009||Oct 6, 2015||Broadcom Corporation||Advanced processor with fast messaging network technology|
|US9264380||May 7, 2011||Feb 16, 2016||Broadcom Corporation||Method and apparatus for implementing cache coherency of a processor|
|US20050027793 *||Aug 31, 2004||Feb 3, 2005||Hass David T.||Advanced processor with mechanism for packet distribution at high line rate|
|US20050033832 *||Aug 31, 2004||Feb 10, 2005||David T. Hass||Advanced processor with use of bridges on a data movement ring for optimal redirection of memory and I/O traffic|
|US20050033889 *||Aug 31, 2004||Feb 10, 2005||Hass David T.||Advanced processor with interrupt delivery mechanism for multi-threaded multi-CPU system on a chip|
|US20050039182 *||Aug 14, 2003||Feb 17, 2005||Hooper Donald F.||Phasing for a multi-threaded network processor|
|US20050041651 *||Aug 31, 2004||Feb 24, 2005||Hass David T.||Advanced processor with mechanism for fast packet queuing operations|
|US20050041666 *||Aug 31, 2004||Feb 24, 2005||Hass David T.||Advanced processor with mechanism for enforcing ordering between information sent on two independent networks|
|US20050044308 *||Aug 31, 2004||Feb 24, 2005||Abbas Rashid||Advanced processor with interfacing messaging network to a CPU|
|US20050044324 *||Aug 31, 2004||Feb 24, 2005||Abbas Rashid||Advanced processor with mechanism for maximizing resource usage in an in-order pipeline with multiple threads|
|US20050055503 *||Jul 23, 2004||Mar 10, 2005||Hass David T.||Advanced processor with cache coherency|
|US20050055540 *||Jul 23, 2004||Mar 10, 2005||Hass David T.||Advanced processor scheduling in a multithreaded system|
|US20050060458 *||Jul 8, 2004||Mar 17, 2005||Yu In-Chul||Method and apparatus for sharing a device among multiple CPU systems|
|US20060095634 *||Nov 1, 2004||May 4, 2006||Meyer Michael J||Method and apparatus for round robin resource arbitration with a fast request to grant response|
|US20070162705 *||Jan 19, 2007||Jul 12, 2007||Adaptec, Inc.||Sequencer Cache and Method for Operating the Same|
|US20070204130 *||Feb 8, 2007||Aug 30, 2007||Raza Microelectronics, Inc.||Advanced processor translation lookaside buffer management in a multithreaded system|
|US20080062927 *||Jul 31, 2007||Mar 13, 2008||Raza Microelectronics, Inc.||Delegating Network Processor Operations to Star Topology Serial Bus Interfaces|
|US20090055496 *||Oct 30, 2008||Feb 26, 2009||Gaurav Garg||Advanced processor with credit based scheme for optimal packet flow in a multi-processor system on a chip|
|US20090201935 *||Feb 8, 2008||Aug 13, 2009||Hass David T||System and method for parsing and allocating a plurality of packets to processor core threads|
|US20100077150 *||Nov 30, 2009||Mar 25, 2010||Rmi Corporation||Advanced processor with cache coherency|
|US20100318703 *||Jun 14, 2010||Dec 16, 2010||Netlogic Microsystems, Inc.||Delegating network processor operations to star topology serial bus interfaces|
|U.S. Classification||710/200, 712/E09.053, 718/102|
|International Classification||G06F9/46, H04L7/00, G06F9/38, G06F9/48|
|Cooperative Classification||G06F9/485, G06F9/3851|
|European Classification||G06F9/38E4, G06F9/48C4P|
|Dec 18, 2003||AS||Assignment|
Owner name: INTEL CORPORATION, CALIFORNIA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:MCALPINE, GARY L.;MINTURN, DAVE B.;REEL/FRAME:014832/0629
Effective date: 20031217
|Jul 15, 2009||FPAY||Fee payment|
Year of fee payment: 4
|Mar 13, 2013||FPAY||Fee payment|
Year of fee payment: 8